Quantum Opt ML DMDD

Quantum Optimization, Machine Learning and Drug and Material Design

The scaling of computing and communication has led to a large wealth creation and lifted large populations from poverty to middle class but can no longer be continued due to our running into physical limits. The use of multi-processing and of many accelerators produces diminishing returns. In particular the use of multi-processing lowers the compute time by computing in parallel but incurs communication time cost to communicate an increasing number of intermediate results from producers to consumers. Quantum computing, by making available new resources, is seen as able to provide an exponential advantage in complexity of computation performed and/or amount of computational resources required.

Wth COVID in recent past, likes of Moderna and Pfizer have become, if they

were not already, household names. The amount of good they do and their

market valuation is often in direct proportion to the number of 2, 5 and now

even 10 billion dollars a year revenue drugs in their portfolio. As broad cures

for diseases such as Cancer are yet to be found and as just as SAARS was

followed by COVID, new pandemics will follow COVID and will need vaccines and therapies, there is room for many block-buster drugs.

While medicines are only used by unwell people, materials are used by un-

well and well people and also industries and governments and can potentially

produce revenues of 20, 50 or even 100 billion dollars a year! Some example

materials are those used in making of batteries. The materials give a battery a

certain capacity, a certain weight and a certain probability of explosion. Can

materials be found that quadruple the capacity and reduce the weight and

probability of explosion to a quarter?

Today with help of computing including learning-aided-computing (‘AI’) we

can already map properties desired in a drug or a material to candidate molec-

ular structures and even find paths in a reaction network from nodes corre-

sponding to readily found molecules to the desired molecule. However the

difficulty lies in the fact that the number of candidate molecules can be very

large and making and testing them in the laboratory can takes 1, 2, 5 or even

10 years! The solution to the basic problem as famously prescribed by Feyn-

man is simulation of Quantum Chemistry using Quantum Computers and, in

this particular case, to replace making-and-testing in the laboratory with such

simulations.

However efficiently performing using a Quantum Computer a task hard for

classical computers requires an appropriate quantum algorithm. While quan-

tum algorithms to efficiently simulate the evolution of a system of many parti-

cles exist, those for predicting the ground state of the Hamiltonian describing

a system of many particles has to resort to heuristics and the latter problem

belongs in the complexity class QMA, the quantum analog of NP. It further turns out that the latter problem can be solved by learning from a polynomial

number of example problem-solution pairs.

It is now well-established that learning for performing and performing learn-

ing aided tasks employing classical computing are done cost-performance ef-

ficiently if done employing hardware acceleration. The plentiful availability

of hardware acceleration for Deep Neural Networks (DNN) (1) has encour-

aged their use as sub-algorithms in larger - Reinforcement Learning (RL) (2),

Generative-Adversarial Networks - algorithms. RL and GAN (and use of DNN

by the RL and GANs for policy and value function approximation) have estab-

lished themselves as advantageous means of performing tasks autonomously

and of performing synthesis tasks. DNN learning is susceptible to inefficien-

cies and failures including vanishing/exploding gradients and local minima.

The two primary means of achieving efficiency in DNN inference - sparsity

and parallelism - conflict.

The simplest Quantum Algorithm - the Deutsch Jozsa - exploits the quantum

resources of Superposition to evaluate all alternatives in parallel (referred to as

in superposition) and the resource Interference to (interfere away all-but and)

be left with the optimal altaernative. Most sophisticated recent algorithm, that

for solving unstructured NP search problem, takes a problem - decoding a list-

recoverable code - known to be classically hard and presents a polynomial

quantum algorithm for the problem.

This implies that given a drug or material design or discovery job, the job

needs to be divided into tasks such that some tasks are best done on a quan-

tum computer, some on a classical computer employing learning and rest on turns out that the latter problem can be solved by learning from a polynomial